Clouds are a very important issue that are not currently handled at all well in models. GCMs are too coarse to directly model clouds, so they are incorporated using parameterisations. Parameterisations ideally come from detailed physical models; but our detailed physical models of clouds don't correspond to reality, so models use parameterisations based on current observations.
There are two problems with this: firstly, and my main bugbear, the present cloud parameterisations are simply wrong with respect to the *spectral response of clouds, particularly in the tropics; secondly, current observations tell us nothing about how clouds may vary under conditions different to those of today. I note Lindzen's Iris hypothesis gets a knock above (and I don't really subscribe to Lindzen's view myself, for other reasons), but his position is really no worse than what is already in the model parameterisations, which boils down to one big SWAG.
Failure to get these fundamental relationships right in a model containing feedback destroys predictive skill in the model. Nice viewpoint on this topic from
Patrick Frank. Again his nuance is slightly different to mine, but he puts his point across very well.
Edited to add: * by spectral response, I am referring to temporal and spatial spectrum, rather than EM spectrum, although our detailed models don't do a good job with the EM spectrum either... but that's another story!
